Research map / Multi-agent systems and games / LLM agents

Resp-Agent: An Agent-Based System for Multimodal Respiratory Sound Generation and Disease Diagnosis

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Resp-Agent, an LLM-orchestrated multi-agent system that pairs a modality-weaving Longformer diagnoser with a flow-matching audio generator (Resp-MLLM) for targeted synthesis of rare respiratory disease sounds, and releases the Resp-229k benchmark.

Problem

Respiratory sound diagnosis suffers from information loss in spectrogram-based representations and scarce, severely class-imbalanced multimodal data, with analysis and generation treated separately.

Approach

A closed-loop planner (Thinker-A2CA) identifies diagnostic weaknesses and requests condition-controlled synthesis from a Qwen3-0.6B-based generator with BEATs style injection and conditional flow matching; the diagnoser weaves EHR text and BEATs audio tokens at the input with sparse anchor global attention in a Longformer.

Key results

• On ICBHI 60-40 split, Resp-Agent reports an ICBHI score of 72.70 (Sp 79.29, Se 66.10), above the prior best of 67.55 reported in the table.
• On Resp-229k Test-CD, Thinker-A2CA synthesis raises the Longformer diagnoser Macro-F1 from 0.212 (no synthesis) to 0.598 at a 50k synthetic budget, versus 0.512 for class-prior and 0.546 for uncertainty-static planners.
• Modality weaving with anchors reaches 0.849 accuracy and 0.212 Macro-F1 on Test-CD, while the audio-only Conformer reaches 0.720 accuracy and 0.1935 Macro-F1.

Surprising findings

• Naive pathology-agnostic augmentation (duplication, pitch/time shift, noise) lowers cross-domain Macro-F1 of the Conformer from 0.1935 to 0.1688.
• Replacing the Conformer with Whisper-Small yields only a small audio-only gain (0.2010 Macro-F1), and text-only Transformers remain far below audio-only models.

Limitations

• Diagnoser ablations and Generator comparisons are reported on a single cross-domain split with severe imbalance, and several tables show inconsistent numbers across sections (e.g., Table 2 and 3 values).
• Clinical summaries are LLM-generated from metadata and audited partly by an LLM verifier, so label and text fidelity are not fully validated.
• Not a certified medical device; no clinical validation performed.

Experimental setup

Datasets: Resp-229k ICBHI SPRSound UK COVID-19 COUGHVID KAUH HF_Lung_V1

Benchmarks: ICBHI 2017 Resp-229k Test-CD

Metrics: accuracy macro-F1 ICBHI score sensitivity specificity FAD cosine similarity

Baselines: Conformer LSTM BERT RoBERTa Whisper-Small c-WaveGAN AudioLDM 2 StableAudio Open Dong et al. AST-based model MVST

Builds on: Longformer BEATs Vocos Qwen3 conditional flow matching

Closest papers by meaning

• Towards Open Respiratory Acoustic Foundation Models: Pretraining and Benchmarking (NeurIPS 2024 · similarity 0.86)
• MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning (ICLR 2026 · similarity 0.82)
• MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments (NeurIPS 2026 · similarity 0.82)
• MedAgent-Pro: Towards Evidence-based Multi-modal Medical Diagnosis via Reasoning Agentic Workflow (ICLR 2026 · similarity 0.82)
• MedVR: Annotation-Free Medical Visual Reasoning via Agentic Reinforcement Learning (ICLR 2026 · similarity 0.81)
• 3DMedAgent: Unified Perception-to-Understanding for 3D Medical Analysis (ICML 2026 · similarity 0.81)